The Third Way

Most discussions about AI safety conflate two different things: human presence and human intent. But you can have one without the other.

You can be present without owning the goal. You watch Netflix for hours but never decided to see another true crime documentary. The algorithm did. You can be absent while owning the goal. You scheduled a bank transfer to pay rent on the first of each month. You might be asleep when it runs. But your intent executes exactly as specified.

Presence and ownership vary independently. They’re separate dials.

This matters because most debates about AI assume one dimension: how much control humans have. More control, safer but slower. Less control, faster but riskier. The whole conversation is just arguing over where to put the slider.

But there’s a second dimension: who owns the goal?

Plot presence on one axis and intent ownership on the other. You get four quadrants.

Upper left: present, human-owned intent. Traditional software. You’re there, in control, deciding everything. Safe because you’re watching. Doesn’t scale because you’re watching.

Lower left: absent, human-owned intent. This is the one everyone misses. You’re not there, but your goals are. The system pursues what you defined, not what it decides.

Upper right: present, system-owned intent. Performative safety at its worst: you’re watching an AI pursue its own goals. Your presence changes nothing. You’re a spectator pretending to be in control.

Lower right: the danger zone. Nobody’s watching and the system picks its own targets. Every AI doomer scenario lives here.

Most people think the choice is between upper left (safe but doesn’t scale) and lower right (scales but dangerous). But there’s a third option: lower left.

What if you could be absent while your intent stays fixed? Not just present. Fixed. The system can’t change what success looks like. Can’t redefine the goal. Can’t decide that actually, this other thing matters more. The system owns execution. You own intent. That boundary doesn’t move.

This breaks the assumed tradeoff. Safety doesn’t require your presence. It requires preserved intent. Different things. Fix intent correctly and you can scale outcomes without scaling attention.

Most AI agents end up in the wrong quadrant, and the problem is subtle. Agents aren’t malicious. They just infer goals instead of receiving them.

You say “help me with my email.” The agent interprets this. What does “help” mean? Which emails matter? How should it prioritize? The agent answers these questions itself. That’s goal inference: the agent generates intent instead of following it.

At first this feels convenient. The agent figures out what you probably want. No need to be precise. Then over time, something shifts. The agent’s interpretation drifts from your actual goals. Small misalignments compound. Eventually you’re watching it pursue something you didn’t ask for.

This is how you slip into the danger zone. Not all at once. Just a thousand small inferences, each one barely noticeable.

The solution is that intent should be explicit. Written down. Authored by humans. The system shouldn’t guess at what you want or try to learn it from your behavior. What does success look like? What is never acceptable? When should the system stop and ask? Write these things down.

This is work. Most people don’t want to do it. “Help me with my email” is easier. But that convenience pulls you toward the danger zone. Every shortcut in intent specification is a step toward system-owned goals.

The people who thrive with AI will be the ones who get good at saying what they actually want. Not because clarity is virtuous, but because it’s the only way to stay in the safe quadrant as your presence drops to zero.

There’s one more piece that makes this work. You can’t watch the system. You’re not there. How do you know when something goes wrong? The system tells you. Confidence drops? It escalates. Context unexpected? It escalates. Constraints at risk? It escalates. Never fails silently.

When it’s working, you hear nothing. When it needs you, it tells you. This flips the usual dynamic. Silence stops being suspicious and becomes information. Not hearing from the system means things are fine. Most systems fail silently all the time, which is why you feel you have to watch them. Design escalation correctly and silence becomes meaningful. You can leave.

The “always supervise” camp thinks safety requires presence. That’s the wrong thing to optimize for. Safety requires fixed intent, not eyeballs. The “autonomy” camp assumes AI should own goals. They want the system to figure out what’s valuable. That’s what makes autonomy dangerous. System-owned intent is the danger zone by definition.

The third way threads between them. You get low presence, which the autonomy camp wants. You keep human-owned intent, which the supervision camp wants. Both benefits, neither cost.

And this isn’t centrism. It’s not “a little supervision, a little autonomy, somewhere in the middle.” It’s a different axis entirely. Supervision and autonomy both accept the tradeoff between safety and scale. They just land in different places. The third way rejects the tradeoff. You can have both if you think about the problem differently. Separate presence from ownership. That’s not splitting the difference. That’s reframing the question.

A lot of AI development is pointed in the wrong direction. We’re building agents that infer goals. The whole pitch is that they figure out what you want. That’s exactly the problem. What we need are systems that lock in goals you’ve defined and refuse to budge from them. Less convenient upfront. Way safer over time.

The measure of a good AI system isn’t how little you have to tell it. It’s how completely you can leave it.

I suspect the most valuable skill in an AI-heavy world won’t be prompt engineering. It won’t be knowing which tools to use or how to chain them together. It will be intent specification: being able to say, precisely and completely, what success looks like, what constraints can’t be violated, what tradeoffs you’re willing to make.

This is harder than it sounds. Most people don’t actually know what they want. They’ve never had to be explicit about their values, because they’ve always been there to course-correct in the moment. But intent specification is the skill that unlocks the third way. Get good at it and you can delegate everything else. Can’t do it? You’re stuck watching forever, or slowly drifting toward the danger zone.

We’re at a fork. One path leads to agents that do more and more on your behalf, inferring your goals, making assumptions, gradually taking ownership of intent. That path ends in the danger zone. The other path leads to systems that do more and more on your behalf, but only within goals you’ve explicitly defined. That path ends in actual freedom: you can leave, and when you come back, the things you cared about got done.

Both involve powerful AI. Both reduce human effort. The difference is who owns the goal.